src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRF_VdwLJ_GeomW4W4_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse2_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_sse2_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_sse2_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  91     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  93     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  95     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  97     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 115     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->ic->epsfac);
 132     charge           = mdatoms->chargeA;
 133     krf              = _mm_set1_ps(fr->ic->k_rf);
 134     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 135     crf              = _mm_set1_ps(fr->ic->c_rf);
 136     nvdwtype         = fr->ntype;
 137     vdwparam         = fr->nbfp;
 138     vdwtype          = mdatoms->typeA;
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 143     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 144     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq1              = _mm_set1_ps(charge[inr+1]);
 148     jq2              = _mm_set1_ps(charge[inr+2]);
 149     jq3              = _mm_set1_ps(charge[inr+3]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq11             = _mm_mul_ps(iq1,jq1);
 154     qq12             = _mm_mul_ps(iq1,jq2);
 155     qq13             = _mm_mul_ps(iq1,jq3);
 156     qq21             = _mm_mul_ps(iq2,jq1);
 157     qq22             = _mm_mul_ps(iq2,jq2);
 158     qq23             = _mm_mul_ps(iq2,jq3);
 159     qq31             = _mm_mul_ps(iq3,jq1);
 160     qq32             = _mm_mul_ps(iq3,jq2);
 161     qq33             = _mm_mul_ps(iq3,jq3);
 162
 163     /* Avoid stupid compiler warnings */
 164     jnrA = jnrB = jnrC = jnrD = 0;
 165     j_coord_offsetA = 0;
 166     j_coord_offsetB = 0;
 167     j_coord_offsetC = 0;
 168     j_coord_offsetD = 0;
 169
 170     outeriter        = 0;
 171     inneriter        = 0;
 172
 173     for(iidx=0;iidx<4*DIM;iidx++)
 174     {
 175         scratch[iidx] = 0.0;
 176     }
 177
 178     /* Start outer loop over neighborlists */
 179     for(iidx=0; iidx<nri; iidx++)
 180     {
 181         /* Load shift vector for this list */
 182         i_shift_offset   = DIM*shiftidx[iidx];
 183
 184         /* Load limits for loop over neighbors */
 185         j_index_start    = jindex[iidx];
 186         j_index_end      = jindex[iidx+1];
 187
 188         /* Get outer coordinate index */
 189         inr              = iinr[iidx];
 190         i_coord_offset   = DIM*inr;
 191
 192         /* Load i particle coords and add shift vector */
 193         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 194                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 195
 196         fix0             = _mm_setzero_ps();
 197         fiy0             = _mm_setzero_ps();
 198         fiz0             = _mm_setzero_ps();
 199         fix1             = _mm_setzero_ps();
 200         fiy1             = _mm_setzero_ps();
 201         fiz1             = _mm_setzero_ps();
 202         fix2             = _mm_setzero_ps();
 203         fiy2             = _mm_setzero_ps();
 204         fiz2             = _mm_setzero_ps();
 205         fix3             = _mm_setzero_ps();
 206         fiy3             = _mm_setzero_ps();
 207         fiz3             = _mm_setzero_ps();
 208
 209         /* Reset potential sums */
 210         velecsum         = _mm_setzero_ps();
 211         vvdwsum          = _mm_setzero_ps();
 212
 213         /* Start inner kernel loop */
 214         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 215         {
 216
 217             /* Get j neighbor index, and coordinate index */
 218             jnrA             = jjnr[jidx];
 219             jnrB             = jjnr[jidx+1];
 220             jnrC             = jjnr[jidx+2];
 221             jnrD             = jjnr[jidx+3];
 222             j_coord_offsetA  = DIM*jnrA;
 223             j_coord_offsetB  = DIM*jnrB;
 224             j_coord_offsetC  = DIM*jnrC;
 225             j_coord_offsetD  = DIM*jnrD;
 226
 227             /* load j atom coordinates */
 228             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 229                                               x+j_coord_offsetC,x+j_coord_offsetD,
 230                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 231                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 232
 233             /* Calculate displacement vector */
 234             dx00             = _mm_sub_ps(ix0,jx0);
 235             dy00             = _mm_sub_ps(iy0,jy0);
 236             dz00             = _mm_sub_ps(iz0,jz0);
 237             dx11             = _mm_sub_ps(ix1,jx1);
 238             dy11             = _mm_sub_ps(iy1,jy1);
 239             dz11             = _mm_sub_ps(iz1,jz1);
 240             dx12             = _mm_sub_ps(ix1,jx2);
 241             dy12             = _mm_sub_ps(iy1,jy2);
 242             dz12             = _mm_sub_ps(iz1,jz2);
 243             dx13             = _mm_sub_ps(ix1,jx3);
 244             dy13             = _mm_sub_ps(iy1,jy3);
 245             dz13             = _mm_sub_ps(iz1,jz3);
 246             dx21             = _mm_sub_ps(ix2,jx1);
 247             dy21             = _mm_sub_ps(iy2,jy1);
 248             dz21             = _mm_sub_ps(iz2,jz1);
 249             dx22             = _mm_sub_ps(ix2,jx2);
 250             dy22             = _mm_sub_ps(iy2,jy2);
 251             dz22             = _mm_sub_ps(iz2,jz2);
 252             dx23             = _mm_sub_ps(ix2,jx3);
 253             dy23             = _mm_sub_ps(iy2,jy3);
 254             dz23             = _mm_sub_ps(iz2,jz3);
 255             dx31             = _mm_sub_ps(ix3,jx1);
 256             dy31             = _mm_sub_ps(iy3,jy1);
 257             dz31             = _mm_sub_ps(iz3,jz1);
 258             dx32             = _mm_sub_ps(ix3,jx2);
 259             dy32             = _mm_sub_ps(iy3,jy2);
 260             dz32             = _mm_sub_ps(iz3,jz2);
 261             dx33             = _mm_sub_ps(ix3,jx3);
 262             dy33             = _mm_sub_ps(iy3,jy3);
 263             dz33             = _mm_sub_ps(iz3,jz3);
 264
 265             /* Calculate squared distance and things based on it */
 266             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 267             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 268             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 269             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 270             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 271             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 272             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 273             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 274             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 275             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 276
 277             rinv11           = sse2_invsqrt_f(rsq11);
 278             rinv12           = sse2_invsqrt_f(rsq12);
 279             rinv13           = sse2_invsqrt_f(rsq13);
 280             rinv21           = sse2_invsqrt_f(rsq21);
 281             rinv22           = sse2_invsqrt_f(rsq22);
 282             rinv23           = sse2_invsqrt_f(rsq23);
 283             rinv31           = sse2_invsqrt_f(rsq31);
 284             rinv32           = sse2_invsqrt_f(rsq32);
 285             rinv33           = sse2_invsqrt_f(rsq33);
 286
 287             rinvsq00         = sse2_inv_f(rsq00);
 288             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 289             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 290             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 291             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 292             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 293             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 294             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 295             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 296             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 297
 298             fjx0             = _mm_setzero_ps();
 299             fjy0             = _mm_setzero_ps();
 300             fjz0             = _mm_setzero_ps();
 301             fjx1             = _mm_setzero_ps();
 302             fjy1             = _mm_setzero_ps();
 303             fjz1             = _mm_setzero_ps();
 304             fjx2             = _mm_setzero_ps();
 305             fjy2             = _mm_setzero_ps();
 306             fjz2             = _mm_setzero_ps();
 307             fjx3             = _mm_setzero_ps();
 308             fjy3             = _mm_setzero_ps();
 309             fjz3             = _mm_setzero_ps();
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* LENNARD-JONES DISPERSION/REPULSION */
 316
 317             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 318             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 319             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 320             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 321             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 322
 323             /* Update potential sum for this i atom from the interaction with this j atom. */
 324             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 325
 326             fscal            = fvdw;
 327
 328             /* Calculate temporary vectorial force */
 329             tx               = _mm_mul_ps(fscal,dx00);
 330             ty               = _mm_mul_ps(fscal,dy00);
 331             tz               = _mm_mul_ps(fscal,dz00);
 332
 333             /* Update vectorial force */
 334             fix0             = _mm_add_ps(fix0,tx);
 335             fiy0             = _mm_add_ps(fiy0,ty);
 336             fiz0             = _mm_add_ps(fiz0,tz);
 337
 338             fjx0             = _mm_add_ps(fjx0,tx);
 339             fjy0             = _mm_add_ps(fjy0,ty);
 340             fjz0             = _mm_add_ps(fjz0,tz);
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             /* REACTION-FIELD ELECTROSTATICS */
 347             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 348             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             velecsum         = _mm_add_ps(velecsum,velec);
 352
 353             fscal            = felec;
 354
 355             /* Calculate temporary vectorial force */
 356             tx               = _mm_mul_ps(fscal,dx11);
 357             ty               = _mm_mul_ps(fscal,dy11);
 358             tz               = _mm_mul_ps(fscal,dz11);
 359
 360             /* Update vectorial force */
 361             fix1             = _mm_add_ps(fix1,tx);
 362             fiy1             = _mm_add_ps(fiy1,ty);
 363             fiz1             = _mm_add_ps(fiz1,tz);
 364
 365             fjx1             = _mm_add_ps(fjx1,tx);
 366             fjy1             = _mm_add_ps(fjy1,ty);
 367             fjz1             = _mm_add_ps(fjz1,tz);
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 375             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 376
 377             /* Update potential sum for this i atom from the interaction with this j atom. */
 378             velecsum         = _mm_add_ps(velecsum,velec);
 379
 380             fscal            = felec;
 381
 382             /* Calculate temporary vectorial force */
 383             tx               = _mm_mul_ps(fscal,dx12);
 384             ty               = _mm_mul_ps(fscal,dy12);
 385             tz               = _mm_mul_ps(fscal,dz12);
 386
 387             /* Update vectorial force */
 388             fix1             = _mm_add_ps(fix1,tx);
 389             fiy1             = _mm_add_ps(fiy1,ty);
 390             fiz1             = _mm_add_ps(fiz1,tz);
 391
 392             fjx2             = _mm_add_ps(fjx2,tx);
 393             fjy2             = _mm_add_ps(fjy2,ty);
 394             fjz2             = _mm_add_ps(fjz2,tz);
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             /* REACTION-FIELD ELECTROSTATICS */
 401             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 402             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velecsum         = _mm_add_ps(velecsum,velec);
 406
 407             fscal            = felec;
 408
 409             /* Calculate temporary vectorial force */
 410             tx               = _mm_mul_ps(fscal,dx13);
 411             ty               = _mm_mul_ps(fscal,dy13);
 412             tz               = _mm_mul_ps(fscal,dz13);
 413
 414             /* Update vectorial force */
 415             fix1             = _mm_add_ps(fix1,tx);
 416             fiy1             = _mm_add_ps(fiy1,ty);
 417             fiz1             = _mm_add_ps(fiz1,tz);
 418
 419             fjx3             = _mm_add_ps(fjx3,tx);
 420             fjy3             = _mm_add_ps(fjy3,ty);
 421             fjz3             = _mm_add_ps(fjz3,tz);
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 429             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 430
 431             /* Update potential sum for this i atom from the interaction with this j atom. */
 432             velecsum         = _mm_add_ps(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             /* Calculate temporary vectorial force */
 437             tx               = _mm_mul_ps(fscal,dx21);
 438             ty               = _mm_mul_ps(fscal,dy21);
 439             tz               = _mm_mul_ps(fscal,dz21);
 440
 441             /* Update vectorial force */
 442             fix2             = _mm_add_ps(fix2,tx);
 443             fiy2             = _mm_add_ps(fiy2,ty);
 444             fiz2             = _mm_add_ps(fiz2,tz);
 445
 446             fjx1             = _mm_add_ps(fjx1,tx);
 447             fjy1             = _mm_add_ps(fjy1,ty);
 448             fjz1             = _mm_add_ps(fjz1,tz);
 449
 450             /**************************
 451              * CALCULATE INTERACTIONS *
 452              **************************/
 453
 454             /* REACTION-FIELD ELECTROSTATICS */
 455             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 456             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 457
 458             /* Update potential sum for this i atom from the interaction with this j atom. */
 459             velecsum         = _mm_add_ps(velecsum,velec);
 460
 461             fscal            = felec;
 462
 463             /* Calculate temporary vectorial force */
 464             tx               = _mm_mul_ps(fscal,dx22);
 465             ty               = _mm_mul_ps(fscal,dy22);
 466             tz               = _mm_mul_ps(fscal,dz22);
 467
 468             /* Update vectorial force */
 469             fix2             = _mm_add_ps(fix2,tx);
 470             fiy2             = _mm_add_ps(fiy2,ty);
 471             fiz2             = _mm_add_ps(fiz2,tz);
 472
 473             fjx2             = _mm_add_ps(fjx2,tx);
 474             fjy2             = _mm_add_ps(fjy2,ty);
 475             fjz2             = _mm_add_ps(fjz2,tz);
 476
 477             /**************************
 478              * CALCULATE INTERACTIONS *
 479              **************************/
 480
 481             /* REACTION-FIELD ELECTROSTATICS */
 482             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 483             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 484
 485             /* Update potential sum for this i atom from the interaction with this j atom. */
 486             velecsum         = _mm_add_ps(velecsum,velec);
 487
 488             fscal            = felec;
 489
 490             /* Calculate temporary vectorial force */
 491             tx               = _mm_mul_ps(fscal,dx23);
 492             ty               = _mm_mul_ps(fscal,dy23);
 493             tz               = _mm_mul_ps(fscal,dz23);
 494
 495             /* Update vectorial force */
 496             fix2             = _mm_add_ps(fix2,tx);
 497             fiy2             = _mm_add_ps(fiy2,ty);
 498             fiz2             = _mm_add_ps(fiz2,tz);
 499
 500             fjx3             = _mm_add_ps(fjx3,tx);
 501             fjy3             = _mm_add_ps(fjy3,ty);
 502             fjz3             = _mm_add_ps(fjz3,tz);
 503
 504             /**************************
 505              * CALCULATE INTERACTIONS *
 506              **************************/
 507
 508             /* REACTION-FIELD ELECTROSTATICS */
 509             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 510             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 511
 512             /* Update potential sum for this i atom from the interaction with this j atom. */
 513             velecsum         = _mm_add_ps(velecsum,velec);
 514
 515             fscal            = felec;
 516
 517             /* Calculate temporary vectorial force */
 518             tx               = _mm_mul_ps(fscal,dx31);
 519             ty               = _mm_mul_ps(fscal,dy31);
 520             tz               = _mm_mul_ps(fscal,dz31);
 521
 522             /* Update vectorial force */
 523             fix3             = _mm_add_ps(fix3,tx);
 524             fiy3             = _mm_add_ps(fiy3,ty);
 525             fiz3             = _mm_add_ps(fiz3,tz);
 526
 527             fjx1             = _mm_add_ps(fjx1,tx);
 528             fjy1             = _mm_add_ps(fjy1,ty);
 529             fjz1             = _mm_add_ps(fjz1,tz);
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             /* REACTION-FIELD ELECTROSTATICS */
 536             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 537             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 538
 539             /* Update potential sum for this i atom from the interaction with this j atom. */
 540             velecsum         = _mm_add_ps(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544             /* Calculate temporary vectorial force */
 545             tx               = _mm_mul_ps(fscal,dx32);
 546             ty               = _mm_mul_ps(fscal,dy32);
 547             tz               = _mm_mul_ps(fscal,dz32);
 548
 549             /* Update vectorial force */
 550             fix3             = _mm_add_ps(fix3,tx);
 551             fiy3             = _mm_add_ps(fiy3,ty);
 552             fiz3             = _mm_add_ps(fiz3,tz);
 553
 554             fjx2             = _mm_add_ps(fjx2,tx);
 555             fjy2             = _mm_add_ps(fjy2,ty);
 556             fjz2             = _mm_add_ps(fjz2,tz);
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             /* REACTION-FIELD ELECTROSTATICS */
 563             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 564             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 565
 566             /* Update potential sum for this i atom from the interaction with this j atom. */
 567             velecsum         = _mm_add_ps(velecsum,velec);
 568
 569             fscal            = felec;
 570
 571             /* Calculate temporary vectorial force */
 572             tx               = _mm_mul_ps(fscal,dx33);
 573             ty               = _mm_mul_ps(fscal,dy33);
 574             tz               = _mm_mul_ps(fscal,dz33);
 575
 576             /* Update vectorial force */
 577             fix3             = _mm_add_ps(fix3,tx);
 578             fiy3             = _mm_add_ps(fiy3,ty);
 579             fiz3             = _mm_add_ps(fiz3,tz);
 580
 581             fjx3             = _mm_add_ps(fjx3,tx);
 582             fjy3             = _mm_add_ps(fjy3,ty);
 583             fjz3             = _mm_add_ps(fjz3,tz);
 584
 585             fjptrA             = f+j_coord_offsetA;
 586             fjptrB             = f+j_coord_offsetB;
 587             fjptrC             = f+j_coord_offsetC;
 588             fjptrD             = f+j_coord_offsetD;
 589
 590             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 591                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 592                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 593
 594             /* Inner loop uses 323 flops */
 595         }
 596
 597         if(jidx<j_index_end)
 598         {
 599
 600             /* Get j neighbor index, and coordinate index */
 601             jnrlistA         = jjnr[jidx];
 602             jnrlistB         = jjnr[jidx+1];
 603             jnrlistC         = jjnr[jidx+2];
 604             jnrlistD         = jjnr[jidx+3];
 605             /* Sign of each element will be negative for non-real atoms.
 606              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 607              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 608              */
 609             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 610             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 611             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 612             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 613             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 614             j_coord_offsetA  = DIM*jnrA;
 615             j_coord_offsetB  = DIM*jnrB;
 616             j_coord_offsetC  = DIM*jnrC;
 617             j_coord_offsetD  = DIM*jnrD;
 618
 619             /* load j atom coordinates */
 620             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 621                                               x+j_coord_offsetC,x+j_coord_offsetD,
 622                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 623                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 624
 625             /* Calculate displacement vector */
 626             dx00             = _mm_sub_ps(ix0,jx0);
 627             dy00             = _mm_sub_ps(iy0,jy0);
 628             dz00             = _mm_sub_ps(iz0,jz0);
 629             dx11             = _mm_sub_ps(ix1,jx1);
 630             dy11             = _mm_sub_ps(iy1,jy1);
 631             dz11             = _mm_sub_ps(iz1,jz1);
 632             dx12             = _mm_sub_ps(ix1,jx2);
 633             dy12             = _mm_sub_ps(iy1,jy2);
 634             dz12             = _mm_sub_ps(iz1,jz2);
 635             dx13             = _mm_sub_ps(ix1,jx3);
 636             dy13             = _mm_sub_ps(iy1,jy3);
 637             dz13             = _mm_sub_ps(iz1,jz3);
 638             dx21             = _mm_sub_ps(ix2,jx1);
 639             dy21             = _mm_sub_ps(iy2,jy1);
 640             dz21             = _mm_sub_ps(iz2,jz1);
 641             dx22             = _mm_sub_ps(ix2,jx2);
 642             dy22             = _mm_sub_ps(iy2,jy2);
 643             dz22             = _mm_sub_ps(iz2,jz2);
 644             dx23             = _mm_sub_ps(ix2,jx3);
 645             dy23             = _mm_sub_ps(iy2,jy3);
 646             dz23             = _mm_sub_ps(iz2,jz3);
 647             dx31             = _mm_sub_ps(ix3,jx1);
 648             dy31             = _mm_sub_ps(iy3,jy1);
 649             dz31             = _mm_sub_ps(iz3,jz1);
 650             dx32             = _mm_sub_ps(ix3,jx2);
 651             dy32             = _mm_sub_ps(iy3,jy2);
 652             dz32             = _mm_sub_ps(iz3,jz2);
 653             dx33             = _mm_sub_ps(ix3,jx3);
 654             dy33             = _mm_sub_ps(iy3,jy3);
 655             dz33             = _mm_sub_ps(iz3,jz3);
 656
 657             /* Calculate squared distance and things based on it */
 658             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 659             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 660             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 661             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 662             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 663             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 664             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 665             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 666             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 667             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 668
 669             rinv11           = sse2_invsqrt_f(rsq11);
 670             rinv12           = sse2_invsqrt_f(rsq12);
 671             rinv13           = sse2_invsqrt_f(rsq13);
 672             rinv21           = sse2_invsqrt_f(rsq21);
 673             rinv22           = sse2_invsqrt_f(rsq22);
 674             rinv23           = sse2_invsqrt_f(rsq23);
 675             rinv31           = sse2_invsqrt_f(rsq31);
 676             rinv32           = sse2_invsqrt_f(rsq32);
 677             rinv33           = sse2_invsqrt_f(rsq33);
 678
 679             rinvsq00         = sse2_inv_f(rsq00);
 680             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 681             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 682             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 683             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 684             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 685             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 686             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 687             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 688             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 689
 690             fjx0             = _mm_setzero_ps();
 691             fjy0             = _mm_setzero_ps();
 692             fjz0             = _mm_setzero_ps();
 693             fjx1             = _mm_setzero_ps();
 694             fjy1             = _mm_setzero_ps();
 695             fjz1             = _mm_setzero_ps();
 696             fjx2             = _mm_setzero_ps();
 697             fjy2             = _mm_setzero_ps();
 698             fjz2             = _mm_setzero_ps();
 699             fjx3             = _mm_setzero_ps();
 700             fjy3             = _mm_setzero_ps();
 701             fjz3             = _mm_setzero_ps();
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             /* LENNARD-JONES DISPERSION/REPULSION */
 708
 709             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 710             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 711             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 712             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 713             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 714
 715             /* Update potential sum for this i atom from the interaction with this j atom. */
 716             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 717             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 718
 719             fscal            = fvdw;
 720
 721             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm_mul_ps(fscal,dx00);
 725             ty               = _mm_mul_ps(fscal,dy00);
 726             tz               = _mm_mul_ps(fscal,dz00);
 727
 728             /* Update vectorial force */
 729             fix0             = _mm_add_ps(fix0,tx);
 730             fiy0             = _mm_add_ps(fiy0,ty);
 731             fiz0             = _mm_add_ps(fiz0,tz);
 732
 733             fjx0             = _mm_add_ps(fjx0,tx);
 734             fjy0             = _mm_add_ps(fjy0,ty);
 735             fjz0             = _mm_add_ps(fjz0,tz);
 736
 737             /**************************
 738              * CALCULATE INTERACTIONS *
 739              **************************/
 740
 741             /* REACTION-FIELD ELECTROSTATICS */
 742             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 743             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 744
 745             /* Update potential sum for this i atom from the interaction with this j atom. */
 746             velec            = _mm_andnot_ps(dummy_mask,velec);
 747             velecsum         = _mm_add_ps(velecsum,velec);
 748
 749             fscal            = felec;
 750
 751             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 752
 753             /* Calculate temporary vectorial force */
 754             tx               = _mm_mul_ps(fscal,dx11);
 755             ty               = _mm_mul_ps(fscal,dy11);
 756             tz               = _mm_mul_ps(fscal,dz11);
 757
 758             /* Update vectorial force */
 759             fix1             = _mm_add_ps(fix1,tx);
 760             fiy1             = _mm_add_ps(fiy1,ty);
 761             fiz1             = _mm_add_ps(fiz1,tz);
 762
 763             fjx1             = _mm_add_ps(fjx1,tx);
 764             fjy1             = _mm_add_ps(fjy1,ty);
 765             fjz1             = _mm_add_ps(fjz1,tz);
 766
 767             /**************************
 768              * CALCULATE INTERACTIONS *
 769              **************************/
 770
 771             /* REACTION-FIELD ELECTROSTATICS */
 772             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 773             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 774
 775             /* Update potential sum for this i atom from the interaction with this j atom. */
 776             velec            = _mm_andnot_ps(dummy_mask,velec);
 777             velecsum         = _mm_add_ps(velecsum,velec);
 778
 779             fscal            = felec;
 780
 781             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 782
 783             /* Calculate temporary vectorial force */
 784             tx               = _mm_mul_ps(fscal,dx12);
 785             ty               = _mm_mul_ps(fscal,dy12);
 786             tz               = _mm_mul_ps(fscal,dz12);
 787
 788             /* Update vectorial force */
 789             fix1             = _mm_add_ps(fix1,tx);
 790             fiy1             = _mm_add_ps(fiy1,ty);
 791             fiz1             = _mm_add_ps(fiz1,tz);
 792
 793             fjx2             = _mm_add_ps(fjx2,tx);
 794             fjy2             = _mm_add_ps(fjy2,ty);
 795             fjz2             = _mm_add_ps(fjz2,tz);
 796
 797             /**************************
 798              * CALCULATE INTERACTIONS *
 799              **************************/
 800
 801             /* REACTION-FIELD ELECTROSTATICS */
 802             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 803             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 804
 805             /* Update potential sum for this i atom from the interaction with this j atom. */
 806             velec            = _mm_andnot_ps(dummy_mask,velec);
 807             velecsum         = _mm_add_ps(velecsum,velec);
 808
 809             fscal            = felec;
 810
 811             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 812
 813             /* Calculate temporary vectorial force */
 814             tx               = _mm_mul_ps(fscal,dx13);
 815             ty               = _mm_mul_ps(fscal,dy13);
 816             tz               = _mm_mul_ps(fscal,dz13);
 817
 818             /* Update vectorial force */
 819             fix1             = _mm_add_ps(fix1,tx);
 820             fiy1             = _mm_add_ps(fiy1,ty);
 821             fiz1             = _mm_add_ps(fiz1,tz);
 822
 823             fjx3             = _mm_add_ps(fjx3,tx);
 824             fjy3             = _mm_add_ps(fjy3,ty);
 825             fjz3             = _mm_add_ps(fjz3,tz);
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 833             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 834
 835             /* Update potential sum for this i atom from the interaction with this j atom. */
 836             velec            = _mm_andnot_ps(dummy_mask,velec);
 837             velecsum         = _mm_add_ps(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 842
 843             /* Calculate temporary vectorial force */
 844             tx               = _mm_mul_ps(fscal,dx21);
 845             ty               = _mm_mul_ps(fscal,dy21);
 846             tz               = _mm_mul_ps(fscal,dz21);
 847
 848             /* Update vectorial force */
 849             fix2             = _mm_add_ps(fix2,tx);
 850             fiy2             = _mm_add_ps(fiy2,ty);
 851             fiz2             = _mm_add_ps(fiz2,tz);
 852
 853             fjx1             = _mm_add_ps(fjx1,tx);
 854             fjy1             = _mm_add_ps(fjy1,ty);
 855             fjz1             = _mm_add_ps(fjz1,tz);
 856
 857             /**************************
 858              * CALCULATE INTERACTIONS *
 859              **************************/
 860
 861             /* REACTION-FIELD ELECTROSTATICS */
 862             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 863             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 864
 865             /* Update potential sum for this i atom from the interaction with this j atom. */
 866             velec            = _mm_andnot_ps(dummy_mask,velec);
 867             velecsum         = _mm_add_ps(velecsum,velec);
 868
 869             fscal            = felec;
 870
 871             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 872
 873             /* Calculate temporary vectorial force */
 874             tx               = _mm_mul_ps(fscal,dx22);
 875             ty               = _mm_mul_ps(fscal,dy22);
 876             tz               = _mm_mul_ps(fscal,dz22);
 877
 878             /* Update vectorial force */
 879             fix2             = _mm_add_ps(fix2,tx);
 880             fiy2             = _mm_add_ps(fiy2,ty);
 881             fiz2             = _mm_add_ps(fiz2,tz);
 882
 883             fjx2             = _mm_add_ps(fjx2,tx);
 884             fjy2             = _mm_add_ps(fjy2,ty);
 885             fjz2             = _mm_add_ps(fjz2,tz);
 886
 887             /**************************
 888              * CALCULATE INTERACTIONS *
 889              **************************/
 890
 891             /* REACTION-FIELD ELECTROSTATICS */
 892             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 893             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 894
 895             /* Update potential sum for this i atom from the interaction with this j atom. */
 896             velec            = _mm_andnot_ps(dummy_mask,velec);
 897             velecsum         = _mm_add_ps(velecsum,velec);
 898
 899             fscal            = felec;
 900
 901             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 902
 903             /* Calculate temporary vectorial force */
 904             tx               = _mm_mul_ps(fscal,dx23);
 905             ty               = _mm_mul_ps(fscal,dy23);
 906             tz               = _mm_mul_ps(fscal,dz23);
 907
 908             /* Update vectorial force */
 909             fix2             = _mm_add_ps(fix2,tx);
 910             fiy2             = _mm_add_ps(fiy2,ty);
 911             fiz2             = _mm_add_ps(fiz2,tz);
 912
 913             fjx3             = _mm_add_ps(fjx3,tx);
 914             fjy3             = _mm_add_ps(fjy3,ty);
 915             fjz3             = _mm_add_ps(fjz3,tz);
 916
 917             /**************************
 918              * CALCULATE INTERACTIONS *
 919              **************************/
 920
 921             /* REACTION-FIELD ELECTROSTATICS */
 922             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 923             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 924
 925             /* Update potential sum for this i atom from the interaction with this j atom. */
 926             velec            = _mm_andnot_ps(dummy_mask,velec);
 927             velecsum         = _mm_add_ps(velecsum,velec);
 928
 929             fscal            = felec;
 930
 931             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 932
 933             /* Calculate temporary vectorial force */
 934             tx               = _mm_mul_ps(fscal,dx31);
 935             ty               = _mm_mul_ps(fscal,dy31);
 936             tz               = _mm_mul_ps(fscal,dz31);
 937
 938             /* Update vectorial force */
 939             fix3             = _mm_add_ps(fix3,tx);
 940             fiy3             = _mm_add_ps(fiy3,ty);
 941             fiz3             = _mm_add_ps(fiz3,tz);
 942
 943             fjx1             = _mm_add_ps(fjx1,tx);
 944             fjy1             = _mm_add_ps(fjy1,ty);
 945             fjz1             = _mm_add_ps(fjz1,tz);
 946
 947             /**************************
 948              * CALCULATE INTERACTIONS *
 949              **************************/
 950
 951             /* REACTION-FIELD ELECTROSTATICS */
 952             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 953             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 954
 955             /* Update potential sum for this i atom from the interaction with this j atom. */
 956             velec            = _mm_andnot_ps(dummy_mask,velec);
 957             velecsum         = _mm_add_ps(velecsum,velec);
 958
 959             fscal            = felec;
 960
 961             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 962
 963             /* Calculate temporary vectorial force */
 964             tx               = _mm_mul_ps(fscal,dx32);
 965             ty               = _mm_mul_ps(fscal,dy32);
 966             tz               = _mm_mul_ps(fscal,dz32);
 967
 968             /* Update vectorial force */
 969             fix3             = _mm_add_ps(fix3,tx);
 970             fiy3             = _mm_add_ps(fiy3,ty);
 971             fiz3             = _mm_add_ps(fiz3,tz);
 972
 973             fjx2             = _mm_add_ps(fjx2,tx);
 974             fjy2             = _mm_add_ps(fjy2,ty);
 975             fjz2             = _mm_add_ps(fjz2,tz);
 976
 977             /**************************
 978              * CALCULATE INTERACTIONS *
 979              **************************/
 980
 981             /* REACTION-FIELD ELECTROSTATICS */
 982             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 983             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 984
 985             /* Update potential sum for this i atom from the interaction with this j atom. */
 986             velec            = _mm_andnot_ps(dummy_mask,velec);
 987             velecsum         = _mm_add_ps(velecsum,velec);
 988
 989             fscal            = felec;
 990
 991             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 992
 993             /* Calculate temporary vectorial force */
 994             tx               = _mm_mul_ps(fscal,dx33);
 995             ty               = _mm_mul_ps(fscal,dy33);
 996             tz               = _mm_mul_ps(fscal,dz33);
 997
 998             /* Update vectorial force */
 999             fix3             = _mm_add_ps(fix3,tx);
1000             fiy3             = _mm_add_ps(fiy3,ty);
1001             fiz3             = _mm_add_ps(fiz3,tz);
1002
1003             fjx3             = _mm_add_ps(fjx3,tx);
1004             fjy3             = _mm_add_ps(fjy3,ty);
1005             fjz3             = _mm_add_ps(fjz3,tz);
1006
1007             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1008             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1009             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1010             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1011
1012             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1013                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1014                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1015
1016             /* Inner loop uses 323 flops */
1017         }
1018
1019         /* End of innermost loop */
1020
1021         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1022                                               f+i_coord_offset,fshift+i_shift_offset);
1023
1024         ggid                        = gid[iidx];
1025         /* Update potential energies */
1026         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1027         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1028
1029         /* Increment number of inner iterations */
1030         inneriter                  += j_index_end - j_index_start;
1031
1032         /* Outer loop uses 26 flops */
1033     }
1034
1035     /* Increment number of outer iterations */
1036     outeriter        += nri;
1037
1038     /* Update outer/inner flops */
1039
1040     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*323);
1041 }
1042 /*
1043  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_single
1044  * Electrostatics interaction: ReactionField
1045  * VdW interaction:            LennardJones
1046  * Geometry:                   Water4-Water4
1047  * Calculate force/pot:        Force
1048  */
1049 void
1050 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_single
1051                     (t_nblist                    * gmx_restrict       nlist,
1052                      rvec                        * gmx_restrict          xx,
1053                      rvec                        * gmx_restrict          ff,
1054                      struct t_forcerec           * gmx_restrict          fr,
1055                      t_mdatoms                   * gmx_restrict     mdatoms,
1056                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1057                      t_nrnb                      * gmx_restrict        nrnb)
1058 {
1059     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1060      * just 0 for non-waters.
1061      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1062      * jnr indices corresponding to data put in the four positions in the SIMD register.
1063      */
1064     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1065     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1066     int              jnrA,jnrB,jnrC,jnrD;
1067     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1068     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1069     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1070     real             rcutoff_scalar;
1071     real             *shiftvec,*fshift,*x,*f;
1072     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1073     real             scratch[4*DIM];
1074     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1075     int              vdwioffset0;
1076     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1077     int              vdwioffset1;
1078     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1079     int              vdwioffset2;
1080     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1081     int              vdwioffset3;
1082     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1083     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1084     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1085     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1086     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1087     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1088     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1089     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1090     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1091     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1092     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1093     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1094     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1095     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1096     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1097     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1098     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1099     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1100     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1101     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1102     real             *charge;
1103     int              nvdwtype;
1104     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1105     int              *vdwtype;
1106     real             *vdwparam;
1107     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1108     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1109     __m128           dummy_mask,cutoff_mask;
1110     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1111     __m128           one     = _mm_set1_ps(1.0);
1112     __m128           two     = _mm_set1_ps(2.0);
1113     x                = xx[0];
1114     f                = ff[0];
1115
1116     nri              = nlist->nri;
1117     iinr             = nlist->iinr;
1118     jindex           = nlist->jindex;
1119     jjnr             = nlist->jjnr;
1120     shiftidx         = nlist->shift;
1121     gid              = nlist->gid;
1122     shiftvec         = fr->shift_vec[0];
1123     fshift           = fr->fshift[0];
1124     facel            = _mm_set1_ps(fr->ic->epsfac);
1125     charge           = mdatoms->chargeA;
1126     krf              = _mm_set1_ps(fr->ic->k_rf);
1127     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1128     crf              = _mm_set1_ps(fr->ic->c_rf);
1129     nvdwtype         = fr->ntype;
1130     vdwparam         = fr->nbfp;
1131     vdwtype          = mdatoms->typeA;
1132
1133     /* Setup water-specific parameters */
1134     inr              = nlist->iinr[0];
1135     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1136     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1137     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1138     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1139
1140     jq1              = _mm_set1_ps(charge[inr+1]);
1141     jq2              = _mm_set1_ps(charge[inr+2]);
1142     jq3              = _mm_set1_ps(charge[inr+3]);
1143     vdwjidx0A        = 2*vdwtype[inr+0];
1144     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1145     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1146     qq11             = _mm_mul_ps(iq1,jq1);
1147     qq12             = _mm_mul_ps(iq1,jq2);
1148     qq13             = _mm_mul_ps(iq1,jq3);
1149     qq21             = _mm_mul_ps(iq2,jq1);
1150     qq22             = _mm_mul_ps(iq2,jq2);
1151     qq23             = _mm_mul_ps(iq2,jq3);
1152     qq31             = _mm_mul_ps(iq3,jq1);
1153     qq32             = _mm_mul_ps(iq3,jq2);
1154     qq33             = _mm_mul_ps(iq3,jq3);
1155
1156     /* Avoid stupid compiler warnings */
1157     jnrA = jnrB = jnrC = jnrD = 0;
1158     j_coord_offsetA = 0;
1159     j_coord_offsetB = 0;
1160     j_coord_offsetC = 0;
1161     j_coord_offsetD = 0;
1162
1163     outeriter        = 0;
1164     inneriter        = 0;
1165
1166     for(iidx=0;iidx<4*DIM;iidx++)
1167     {
1168         scratch[iidx] = 0.0;
1169     }
1170
1171     /* Start outer loop over neighborlists */
1172     for(iidx=0; iidx<nri; iidx++)
1173     {
1174         /* Load shift vector for this list */
1175         i_shift_offset   = DIM*shiftidx[iidx];
1176
1177         /* Load limits for loop over neighbors */
1178         j_index_start    = jindex[iidx];
1179         j_index_end      = jindex[iidx+1];
1180
1181         /* Get outer coordinate index */
1182         inr              = iinr[iidx];
1183         i_coord_offset   = DIM*inr;
1184
1185         /* Load i particle coords and add shift vector */
1186         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1187                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1188
1189         fix0             = _mm_setzero_ps();
1190         fiy0             = _mm_setzero_ps();
1191         fiz0             = _mm_setzero_ps();
1192         fix1             = _mm_setzero_ps();
1193         fiy1             = _mm_setzero_ps();
1194         fiz1             = _mm_setzero_ps();
1195         fix2             = _mm_setzero_ps();
1196         fiy2             = _mm_setzero_ps();
1197         fiz2             = _mm_setzero_ps();
1198         fix3             = _mm_setzero_ps();
1199         fiy3             = _mm_setzero_ps();
1200         fiz3             = _mm_setzero_ps();
1201
1202         /* Start inner kernel loop */
1203         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1204         {
1205
1206             /* Get j neighbor index, and coordinate index */
1207             jnrA             = jjnr[jidx];
1208             jnrB             = jjnr[jidx+1];
1209             jnrC             = jjnr[jidx+2];
1210             jnrD             = jjnr[jidx+3];
1211             j_coord_offsetA  = DIM*jnrA;
1212             j_coord_offsetB  = DIM*jnrB;
1213             j_coord_offsetC  = DIM*jnrC;
1214             j_coord_offsetD  = DIM*jnrD;
1215
1216             /* load j atom coordinates */
1217             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1218                                               x+j_coord_offsetC,x+j_coord_offsetD,
1219                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1220                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1221
1222             /* Calculate displacement vector */
1223             dx00             = _mm_sub_ps(ix0,jx0);
1224             dy00             = _mm_sub_ps(iy0,jy0);
1225             dz00             = _mm_sub_ps(iz0,jz0);
1226             dx11             = _mm_sub_ps(ix1,jx1);
1227             dy11             = _mm_sub_ps(iy1,jy1);
1228             dz11             = _mm_sub_ps(iz1,jz1);
1229             dx12             = _mm_sub_ps(ix1,jx2);
1230             dy12             = _mm_sub_ps(iy1,jy2);
1231             dz12             = _mm_sub_ps(iz1,jz2);
1232             dx13             = _mm_sub_ps(ix1,jx3);
1233             dy13             = _mm_sub_ps(iy1,jy3);
1234             dz13             = _mm_sub_ps(iz1,jz3);
1235             dx21             = _mm_sub_ps(ix2,jx1);
1236             dy21             = _mm_sub_ps(iy2,jy1);
1237             dz21             = _mm_sub_ps(iz2,jz1);
1238             dx22             = _mm_sub_ps(ix2,jx2);
1239             dy22             = _mm_sub_ps(iy2,jy2);
1240             dz22             = _mm_sub_ps(iz2,jz2);
1241             dx23             = _mm_sub_ps(ix2,jx3);
1242             dy23             = _mm_sub_ps(iy2,jy3);
1243             dz23             = _mm_sub_ps(iz2,jz3);
1244             dx31             = _mm_sub_ps(ix3,jx1);
1245             dy31             = _mm_sub_ps(iy3,jy1);
1246             dz31             = _mm_sub_ps(iz3,jz1);
1247             dx32             = _mm_sub_ps(ix3,jx2);
1248             dy32             = _mm_sub_ps(iy3,jy2);
1249             dz32             = _mm_sub_ps(iz3,jz2);
1250             dx33             = _mm_sub_ps(ix3,jx3);
1251             dy33             = _mm_sub_ps(iy3,jy3);
1252             dz33             = _mm_sub_ps(iz3,jz3);
1253
1254             /* Calculate squared distance and things based on it */
1255             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1256             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1257             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1258             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1259             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1260             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1261             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1262             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1263             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1264             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1265
1266             rinv11           = sse2_invsqrt_f(rsq11);
1267             rinv12           = sse2_invsqrt_f(rsq12);
1268             rinv13           = sse2_invsqrt_f(rsq13);
1269             rinv21           = sse2_invsqrt_f(rsq21);
1270             rinv22           = sse2_invsqrt_f(rsq22);
1271             rinv23           = sse2_invsqrt_f(rsq23);
1272             rinv31           = sse2_invsqrt_f(rsq31);
1273             rinv32           = sse2_invsqrt_f(rsq32);
1274             rinv33           = sse2_invsqrt_f(rsq33);
1275
1276             rinvsq00         = sse2_inv_f(rsq00);
1277             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1278             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1279             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1280             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1281             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1282             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1283             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1284             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1285             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1286
1287             fjx0             = _mm_setzero_ps();
1288             fjy0             = _mm_setzero_ps();
1289             fjz0             = _mm_setzero_ps();
1290             fjx1             = _mm_setzero_ps();
1291             fjy1             = _mm_setzero_ps();
1292             fjz1             = _mm_setzero_ps();
1293             fjx2             = _mm_setzero_ps();
1294             fjy2             = _mm_setzero_ps();
1295             fjz2             = _mm_setzero_ps();
1296             fjx3             = _mm_setzero_ps();
1297             fjy3             = _mm_setzero_ps();
1298             fjz3             = _mm_setzero_ps();
1299
1300             /**************************
1301              * CALCULATE INTERACTIONS *
1302              **************************/
1303
1304             /* LENNARD-JONES DISPERSION/REPULSION */
1305
1306             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1307             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1308
1309             fscal            = fvdw;
1310
1311             /* Calculate temporary vectorial force */
1312             tx               = _mm_mul_ps(fscal,dx00);
1313             ty               = _mm_mul_ps(fscal,dy00);
1314             tz               = _mm_mul_ps(fscal,dz00);
1315
1316             /* Update vectorial force */
1317             fix0             = _mm_add_ps(fix0,tx);
1318             fiy0             = _mm_add_ps(fiy0,ty);
1319             fiz0             = _mm_add_ps(fiz0,tz);
1320
1321             fjx0             = _mm_add_ps(fjx0,tx);
1322             fjy0             = _mm_add_ps(fjy0,ty);
1323             fjz0             = _mm_add_ps(fjz0,tz);
1324
1325             /**************************
1326              * CALCULATE INTERACTIONS *
1327              **************************/
1328
1329             /* REACTION-FIELD ELECTROSTATICS */
1330             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1331
1332             fscal            = felec;
1333
1334             /* Calculate temporary vectorial force */
1335             tx               = _mm_mul_ps(fscal,dx11);
1336             ty               = _mm_mul_ps(fscal,dy11);
1337             tz               = _mm_mul_ps(fscal,dz11);
1338
1339             /* Update vectorial force */
1340             fix1             = _mm_add_ps(fix1,tx);
1341             fiy1             = _mm_add_ps(fiy1,ty);
1342             fiz1             = _mm_add_ps(fiz1,tz);
1343
1344             fjx1             = _mm_add_ps(fjx1,tx);
1345             fjy1             = _mm_add_ps(fjy1,ty);
1346             fjz1             = _mm_add_ps(fjz1,tz);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* REACTION-FIELD ELECTROSTATICS */
1353             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1354
1355             fscal            = felec;
1356
1357             /* Calculate temporary vectorial force */
1358             tx               = _mm_mul_ps(fscal,dx12);
1359             ty               = _mm_mul_ps(fscal,dy12);
1360             tz               = _mm_mul_ps(fscal,dz12);
1361
1362             /* Update vectorial force */
1363             fix1             = _mm_add_ps(fix1,tx);
1364             fiy1             = _mm_add_ps(fiy1,ty);
1365             fiz1             = _mm_add_ps(fiz1,tz);
1366
1367             fjx2             = _mm_add_ps(fjx2,tx);
1368             fjy2             = _mm_add_ps(fjy2,ty);
1369             fjz2             = _mm_add_ps(fjz2,tz);
1370
1371             /**************************
1372              * CALCULATE INTERACTIONS *
1373              **************************/
1374
1375             /* REACTION-FIELD ELECTROSTATICS */
1376             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1377
1378             fscal            = felec;
1379
1380             /* Calculate temporary vectorial force */
1381             tx               = _mm_mul_ps(fscal,dx13);
1382             ty               = _mm_mul_ps(fscal,dy13);
1383             tz               = _mm_mul_ps(fscal,dz13);
1384
1385             /* Update vectorial force */
1386             fix1             = _mm_add_ps(fix1,tx);
1387             fiy1             = _mm_add_ps(fiy1,ty);
1388             fiz1             = _mm_add_ps(fiz1,tz);
1389
1390             fjx3             = _mm_add_ps(fjx3,tx);
1391             fjy3             = _mm_add_ps(fjy3,ty);
1392             fjz3             = _mm_add_ps(fjz3,tz);
1393
1394             /**************************
1395              * CALCULATE INTERACTIONS *
1396              **************************/
1397
1398             /* REACTION-FIELD ELECTROSTATICS */
1399             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1400
1401             fscal            = felec;
1402
1403             /* Calculate temporary vectorial force */
1404             tx               = _mm_mul_ps(fscal,dx21);
1405             ty               = _mm_mul_ps(fscal,dy21);
1406             tz               = _mm_mul_ps(fscal,dz21);
1407
1408             /* Update vectorial force */
1409             fix2             = _mm_add_ps(fix2,tx);
1410             fiy2             = _mm_add_ps(fiy2,ty);
1411             fiz2             = _mm_add_ps(fiz2,tz);
1412
1413             fjx1             = _mm_add_ps(fjx1,tx);
1414             fjy1             = _mm_add_ps(fjy1,ty);
1415             fjz1             = _mm_add_ps(fjz1,tz);
1416
1417             /**************************
1418              * CALCULATE INTERACTIONS *
1419              **************************/
1420
1421             /* REACTION-FIELD ELECTROSTATICS */
1422             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1423
1424             fscal            = felec;
1425
1426             /* Calculate temporary vectorial force */
1427             tx               = _mm_mul_ps(fscal,dx22);
1428             ty               = _mm_mul_ps(fscal,dy22);
1429             tz               = _mm_mul_ps(fscal,dz22);
1430
1431             /* Update vectorial force */
1432             fix2             = _mm_add_ps(fix2,tx);
1433             fiy2             = _mm_add_ps(fiy2,ty);
1434             fiz2             = _mm_add_ps(fiz2,tz);
1435
1436             fjx2             = _mm_add_ps(fjx2,tx);
1437             fjy2             = _mm_add_ps(fjy2,ty);
1438             fjz2             = _mm_add_ps(fjz2,tz);
1439
1440             /**************************
1441              * CALCULATE INTERACTIONS *
1442              **************************/
1443
1444             /* REACTION-FIELD ELECTROSTATICS */
1445             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1446
1447             fscal            = felec;
1448
1449             /* Calculate temporary vectorial force */
1450             tx               = _mm_mul_ps(fscal,dx23);
1451             ty               = _mm_mul_ps(fscal,dy23);
1452             tz               = _mm_mul_ps(fscal,dz23);
1453
1454             /* Update vectorial force */
1455             fix2             = _mm_add_ps(fix2,tx);
1456             fiy2             = _mm_add_ps(fiy2,ty);
1457             fiz2             = _mm_add_ps(fiz2,tz);
1458
1459             fjx3             = _mm_add_ps(fjx3,tx);
1460             fjy3             = _mm_add_ps(fjy3,ty);
1461             fjz3             = _mm_add_ps(fjz3,tz);
1462
1463             /**************************
1464              * CALCULATE INTERACTIONS *
1465              **************************/
1466
1467             /* REACTION-FIELD ELECTROSTATICS */
1468             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1469
1470             fscal            = felec;
1471
1472             /* Calculate temporary vectorial force */
1473             tx               = _mm_mul_ps(fscal,dx31);
1474             ty               = _mm_mul_ps(fscal,dy31);
1475             tz               = _mm_mul_ps(fscal,dz31);
1476
1477             /* Update vectorial force */
1478             fix3             = _mm_add_ps(fix3,tx);
1479             fiy3             = _mm_add_ps(fiy3,ty);
1480             fiz3             = _mm_add_ps(fiz3,tz);
1481
1482             fjx1             = _mm_add_ps(fjx1,tx);
1483             fjy1             = _mm_add_ps(fjy1,ty);
1484             fjz1             = _mm_add_ps(fjz1,tz);
1485
1486             /**************************
1487              * CALCULATE INTERACTIONS *
1488              **************************/
1489
1490             /* REACTION-FIELD ELECTROSTATICS */
1491             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1492
1493             fscal            = felec;
1494
1495             /* Calculate temporary vectorial force */
1496             tx               = _mm_mul_ps(fscal,dx32);
1497             ty               = _mm_mul_ps(fscal,dy32);
1498             tz               = _mm_mul_ps(fscal,dz32);
1499
1500             /* Update vectorial force */
1501             fix3             = _mm_add_ps(fix3,tx);
1502             fiy3             = _mm_add_ps(fiy3,ty);
1503             fiz3             = _mm_add_ps(fiz3,tz);
1504
1505             fjx2             = _mm_add_ps(fjx2,tx);
1506             fjy2             = _mm_add_ps(fjy2,ty);
1507             fjz2             = _mm_add_ps(fjz2,tz);
1508
1509             /**************************
1510              * CALCULATE INTERACTIONS *
1511              **************************/
1512
1513             /* REACTION-FIELD ELECTROSTATICS */
1514             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1515
1516             fscal            = felec;
1517
1518             /* Calculate temporary vectorial force */
1519             tx               = _mm_mul_ps(fscal,dx33);
1520             ty               = _mm_mul_ps(fscal,dy33);
1521             tz               = _mm_mul_ps(fscal,dz33);
1522
1523             /* Update vectorial force */
1524             fix3             = _mm_add_ps(fix3,tx);
1525             fiy3             = _mm_add_ps(fiy3,ty);
1526             fiz3             = _mm_add_ps(fiz3,tz);
1527
1528             fjx3             = _mm_add_ps(fjx3,tx);
1529             fjy3             = _mm_add_ps(fjy3,ty);
1530             fjz3             = _mm_add_ps(fjz3,tz);
1531
1532             fjptrA             = f+j_coord_offsetA;
1533             fjptrB             = f+j_coord_offsetB;
1534             fjptrC             = f+j_coord_offsetC;
1535             fjptrD             = f+j_coord_offsetD;
1536
1537             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1538                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1539                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1540
1541             /* Inner loop uses 273 flops */
1542         }
1543
1544         if(jidx<j_index_end)
1545         {
1546
1547             /* Get j neighbor index, and coordinate index */
1548             jnrlistA         = jjnr[jidx];
1549             jnrlistB         = jjnr[jidx+1];
1550             jnrlistC         = jjnr[jidx+2];
1551             jnrlistD         = jjnr[jidx+3];
1552             /* Sign of each element will be negative for non-real atoms.
1553              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1554              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1555              */
1556             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1557             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1558             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1559             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1560             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1561             j_coord_offsetA  = DIM*jnrA;
1562             j_coord_offsetB  = DIM*jnrB;
1563             j_coord_offsetC  = DIM*jnrC;
1564             j_coord_offsetD  = DIM*jnrD;
1565
1566             /* load j atom coordinates */
1567             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1568                                               x+j_coord_offsetC,x+j_coord_offsetD,
1569                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1570                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1571
1572             /* Calculate displacement vector */
1573             dx00             = _mm_sub_ps(ix0,jx0);
1574             dy00             = _mm_sub_ps(iy0,jy0);
1575             dz00             = _mm_sub_ps(iz0,jz0);
1576             dx11             = _mm_sub_ps(ix1,jx1);
1577             dy11             = _mm_sub_ps(iy1,jy1);
1578             dz11             = _mm_sub_ps(iz1,jz1);
1579             dx12             = _mm_sub_ps(ix1,jx2);
1580             dy12             = _mm_sub_ps(iy1,jy2);
1581             dz12             = _mm_sub_ps(iz1,jz2);
1582             dx13             = _mm_sub_ps(ix1,jx3);
1583             dy13             = _mm_sub_ps(iy1,jy3);
1584             dz13             = _mm_sub_ps(iz1,jz3);
1585             dx21             = _mm_sub_ps(ix2,jx1);
1586             dy21             = _mm_sub_ps(iy2,jy1);
1587             dz21             = _mm_sub_ps(iz2,jz1);
1588             dx22             = _mm_sub_ps(ix2,jx2);
1589             dy22             = _mm_sub_ps(iy2,jy2);
1590             dz22             = _mm_sub_ps(iz2,jz2);
1591             dx23             = _mm_sub_ps(ix2,jx3);
1592             dy23             = _mm_sub_ps(iy2,jy3);
1593             dz23             = _mm_sub_ps(iz2,jz3);
1594             dx31             = _mm_sub_ps(ix3,jx1);
1595             dy31             = _mm_sub_ps(iy3,jy1);
1596             dz31             = _mm_sub_ps(iz3,jz1);
1597             dx32             = _mm_sub_ps(ix3,jx2);
1598             dy32             = _mm_sub_ps(iy3,jy2);
1599             dz32             = _mm_sub_ps(iz3,jz2);
1600             dx33             = _mm_sub_ps(ix3,jx3);
1601             dy33             = _mm_sub_ps(iy3,jy3);
1602             dz33             = _mm_sub_ps(iz3,jz3);
1603
1604             /* Calculate squared distance and things based on it */
1605             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1606             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1607             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1608             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1609             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1610             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1611             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1612             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1613             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1614             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1615
1616             rinv11           = sse2_invsqrt_f(rsq11);
1617             rinv12           = sse2_invsqrt_f(rsq12);
1618             rinv13           = sse2_invsqrt_f(rsq13);
1619             rinv21           = sse2_invsqrt_f(rsq21);
1620             rinv22           = sse2_invsqrt_f(rsq22);
1621             rinv23           = sse2_invsqrt_f(rsq23);
1622             rinv31           = sse2_invsqrt_f(rsq31);
1623             rinv32           = sse2_invsqrt_f(rsq32);
1624             rinv33           = sse2_invsqrt_f(rsq33);
1625
1626             rinvsq00         = sse2_inv_f(rsq00);
1627             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1628             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1629             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1630             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1631             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1632             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1633             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1634             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1635             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1636
1637             fjx0             = _mm_setzero_ps();
1638             fjy0             = _mm_setzero_ps();
1639             fjz0             = _mm_setzero_ps();
1640             fjx1             = _mm_setzero_ps();
1641             fjy1             = _mm_setzero_ps();
1642             fjz1             = _mm_setzero_ps();
1643             fjx2             = _mm_setzero_ps();
1644             fjy2             = _mm_setzero_ps();
1645             fjz2             = _mm_setzero_ps();
1646             fjx3             = _mm_setzero_ps();
1647             fjy3             = _mm_setzero_ps();
1648             fjz3             = _mm_setzero_ps();
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             /* LENNARD-JONES DISPERSION/REPULSION */
1655
1656             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1657             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1658
1659             fscal            = fvdw;
1660
1661             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1662
1663             /* Calculate temporary vectorial force */
1664             tx               = _mm_mul_ps(fscal,dx00);
1665             ty               = _mm_mul_ps(fscal,dy00);
1666             tz               = _mm_mul_ps(fscal,dz00);
1667
1668             /* Update vectorial force */
1669             fix0             = _mm_add_ps(fix0,tx);
1670             fiy0             = _mm_add_ps(fiy0,ty);
1671             fiz0             = _mm_add_ps(fiz0,tz);
1672
1673             fjx0             = _mm_add_ps(fjx0,tx);
1674             fjy0             = _mm_add_ps(fjy0,ty);
1675             fjz0             = _mm_add_ps(fjz0,tz);
1676
1677             /**************************
1678              * CALCULATE INTERACTIONS *
1679              **************************/
1680
1681             /* REACTION-FIELD ELECTROSTATICS */
1682             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1683
1684             fscal            = felec;
1685
1686             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1687
1688             /* Calculate temporary vectorial force */
1689             tx               = _mm_mul_ps(fscal,dx11);
1690             ty               = _mm_mul_ps(fscal,dy11);
1691             tz               = _mm_mul_ps(fscal,dz11);
1692
1693             /* Update vectorial force */
1694             fix1             = _mm_add_ps(fix1,tx);
1695             fiy1             = _mm_add_ps(fiy1,ty);
1696             fiz1             = _mm_add_ps(fiz1,tz);
1697
1698             fjx1             = _mm_add_ps(fjx1,tx);
1699             fjy1             = _mm_add_ps(fjy1,ty);
1700             fjz1             = _mm_add_ps(fjz1,tz);
1701
1702             /**************************
1703              * CALCULATE INTERACTIONS *
1704              **************************/
1705
1706             /* REACTION-FIELD ELECTROSTATICS */
1707             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1708
1709             fscal            = felec;
1710
1711             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1712
1713             /* Calculate temporary vectorial force */
1714             tx               = _mm_mul_ps(fscal,dx12);
1715             ty               = _mm_mul_ps(fscal,dy12);
1716             tz               = _mm_mul_ps(fscal,dz12);
1717
1718             /* Update vectorial force */
1719             fix1             = _mm_add_ps(fix1,tx);
1720             fiy1             = _mm_add_ps(fiy1,ty);
1721             fiz1             = _mm_add_ps(fiz1,tz);
1722
1723             fjx2             = _mm_add_ps(fjx2,tx);
1724             fjy2             = _mm_add_ps(fjy2,ty);
1725             fjz2             = _mm_add_ps(fjz2,tz);
1726
1727             /**************************
1728              * CALCULATE INTERACTIONS *
1729              **************************/
1730
1731             /* REACTION-FIELD ELECTROSTATICS */
1732             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1733
1734             fscal            = felec;
1735
1736             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1737
1738             /* Calculate temporary vectorial force */
1739             tx               = _mm_mul_ps(fscal,dx13);
1740             ty               = _mm_mul_ps(fscal,dy13);
1741             tz               = _mm_mul_ps(fscal,dz13);
1742
1743             /* Update vectorial force */
1744             fix1             = _mm_add_ps(fix1,tx);
1745             fiy1             = _mm_add_ps(fiy1,ty);
1746             fiz1             = _mm_add_ps(fiz1,tz);
1747
1748             fjx3             = _mm_add_ps(fjx3,tx);
1749             fjy3             = _mm_add_ps(fjy3,ty);
1750             fjz3             = _mm_add_ps(fjz3,tz);
1751
1752             /**************************
1753              * CALCULATE INTERACTIONS *
1754              **************************/
1755
1756             /* REACTION-FIELD ELECTROSTATICS */
1757             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1758
1759             fscal            = felec;
1760
1761             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1762
1763             /* Calculate temporary vectorial force */
1764             tx               = _mm_mul_ps(fscal,dx21);
1765             ty               = _mm_mul_ps(fscal,dy21);
1766             tz               = _mm_mul_ps(fscal,dz21);
1767
1768             /* Update vectorial force */
1769             fix2             = _mm_add_ps(fix2,tx);
1770             fiy2             = _mm_add_ps(fiy2,ty);
1771             fiz2             = _mm_add_ps(fiz2,tz);
1772
1773             fjx1             = _mm_add_ps(fjx1,tx);
1774             fjy1             = _mm_add_ps(fjy1,ty);
1775             fjz1             = _mm_add_ps(fjz1,tz);
1776
1777             /**************************
1778              * CALCULATE INTERACTIONS *
1779              **************************/
1780
1781             /* REACTION-FIELD ELECTROSTATICS */
1782             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1783
1784             fscal            = felec;
1785
1786             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1787
1788             /* Calculate temporary vectorial force */
1789             tx               = _mm_mul_ps(fscal,dx22);
1790             ty               = _mm_mul_ps(fscal,dy22);
1791             tz               = _mm_mul_ps(fscal,dz22);
1792
1793             /* Update vectorial force */
1794             fix2             = _mm_add_ps(fix2,tx);
1795             fiy2             = _mm_add_ps(fiy2,ty);
1796             fiz2             = _mm_add_ps(fiz2,tz);
1797
1798             fjx2             = _mm_add_ps(fjx2,tx);
1799             fjy2             = _mm_add_ps(fjy2,ty);
1800             fjz2             = _mm_add_ps(fjz2,tz);
1801
1802             /**************************
1803              * CALCULATE INTERACTIONS *
1804              **************************/
1805
1806             /* REACTION-FIELD ELECTROSTATICS */
1807             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1808
1809             fscal            = felec;
1810
1811             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1812
1813             /* Calculate temporary vectorial force */
1814             tx               = _mm_mul_ps(fscal,dx23);
1815             ty               = _mm_mul_ps(fscal,dy23);
1816             tz               = _mm_mul_ps(fscal,dz23);
1817
1818             /* Update vectorial force */
1819             fix2             = _mm_add_ps(fix2,tx);
1820             fiy2             = _mm_add_ps(fiy2,ty);
1821             fiz2             = _mm_add_ps(fiz2,tz);
1822
1823             fjx3             = _mm_add_ps(fjx3,tx);
1824             fjy3             = _mm_add_ps(fjy3,ty);
1825             fjz3             = _mm_add_ps(fjz3,tz);
1826
1827             /**************************
1828              * CALCULATE INTERACTIONS *
1829              **************************/
1830
1831             /* REACTION-FIELD ELECTROSTATICS */
1832             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1833
1834             fscal            = felec;
1835
1836             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1837
1838             /* Calculate temporary vectorial force */
1839             tx               = _mm_mul_ps(fscal,dx31);
1840             ty               = _mm_mul_ps(fscal,dy31);
1841             tz               = _mm_mul_ps(fscal,dz31);
1842
1843             /* Update vectorial force */
1844             fix3             = _mm_add_ps(fix3,tx);
1845             fiy3             = _mm_add_ps(fiy3,ty);
1846             fiz3             = _mm_add_ps(fiz3,tz);
1847
1848             fjx1             = _mm_add_ps(fjx1,tx);
1849             fjy1             = _mm_add_ps(fjy1,ty);
1850             fjz1             = _mm_add_ps(fjz1,tz);
1851
1852             /**************************
1853              * CALCULATE INTERACTIONS *
1854              **************************/
1855
1856             /* REACTION-FIELD ELECTROSTATICS */
1857             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1858
1859             fscal            = felec;
1860
1861             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1862
1863             /* Calculate temporary vectorial force */
1864             tx               = _mm_mul_ps(fscal,dx32);
1865             ty               = _mm_mul_ps(fscal,dy32);
1866             tz               = _mm_mul_ps(fscal,dz32);
1867
1868             /* Update vectorial force */
1869             fix3             = _mm_add_ps(fix3,tx);
1870             fiy3             = _mm_add_ps(fiy3,ty);
1871             fiz3             = _mm_add_ps(fiz3,tz);
1872
1873             fjx2             = _mm_add_ps(fjx2,tx);
1874             fjy2             = _mm_add_ps(fjy2,ty);
1875             fjz2             = _mm_add_ps(fjz2,tz);
1876
1877             /**************************
1878              * CALCULATE INTERACTIONS *
1879              **************************/
1880
1881             /* REACTION-FIELD ELECTROSTATICS */
1882             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1883
1884             fscal            = felec;
1885
1886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1887
1888             /* Calculate temporary vectorial force */
1889             tx               = _mm_mul_ps(fscal,dx33);
1890             ty               = _mm_mul_ps(fscal,dy33);
1891             tz               = _mm_mul_ps(fscal,dz33);
1892
1893             /* Update vectorial force */
1894             fix3             = _mm_add_ps(fix3,tx);
1895             fiy3             = _mm_add_ps(fiy3,ty);
1896             fiz3             = _mm_add_ps(fiz3,tz);
1897
1898             fjx3             = _mm_add_ps(fjx3,tx);
1899             fjy3             = _mm_add_ps(fjy3,ty);
1900             fjz3             = _mm_add_ps(fjz3,tz);
1901
1902             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1903             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1904             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1905             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1906
1907             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1908                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1909                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1910
1911             /* Inner loop uses 273 flops */
1912         }
1913
1914         /* End of innermost loop */
1915
1916         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1917                                               f+i_coord_offset,fshift+i_shift_offset);
1918
1919         /* Increment number of inner iterations */
1920         inneriter                  += j_index_end - j_index_start;
1921
1922         /* Outer loop uses 24 flops */
1923     }
1924
1925     /* Increment number of outer iterations */
1926     outeriter        += nri;
1927
1928     /* Update outer/inner flops */
1929
1930     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1931 }